How to avoid vendor lock-in in AI and data projects
Vendor lock-in is a dependence on a provider that makes switching so expensive or laborious that it does not happen. This page shows where vendor lock-in arises in AI and data projects, what you can clarify before signing a contract, when a commitment to one provider is acceptable, and what a switch costs.
What Vendor Lock-in, the Lock-in Effect, and Switching Costs Mean
Vendor lock-in is a dependence on a provider that makes switching so expensive or laborious that it does not happen. The EU Data Act uses the term "vendor lock-in" in its recitals [1].
The lock-in effect is the impact that high switching costs have on the customer. The more expensive a switch would be, the weaker the incentive to choose a different provider. In AI and data projects, dependence on the provider usually grows with the length of use, because the provider holds more data and more interfaces to the company's own systems are added.
Switching costs are all the costs caused by a change of provider. They include charges from the previous provider, the working time needed to transfer data and models, the work of rebuilding functions, and the fees for the months in which the old and the new solution run side by side. The level of switching costs shows how strongly a company depends on its provider.
Digital sovereignty is the broader term. It covers freedom of choice as a whole, including the questions of where the data is stored and who owns the model. Vendor lock-in concerns one of these questions, namely whether switching providers remains economically feasible.
The four checks for freedom of choice are described on the page Digital Sovereignty in AI for Mid-Sized Companies.Five Areas Where Vendor Lock-in Arises in AI and Data Projects
In AI and data projects, we distinguish five areas in which dependence can arise. For each area, the table states how the dependence comes about and which safeguard you can put in place before signing a contract. You can use the third column as a checklist for an offer.
| Area | How the dependence arises | Safeguard before signing the contract |
|---|---|---|
| Data formats and data export | The data is stored in a format that only the provider's platform can read, or the provider charges a fee for exporting large volumes of data. | Specify in the contract the format in which you will get all of your data back, the deadline for the return, and the cost of the export. |
| Functions that exist on only one platform | Data preparation, training, or delivery uses functions that no other provider offers in the same form. When you switch, these functions have to be rebuilt. | Ask for a list of the parts of the solution that use such functions. Use these functions only where their benefit justifies the effort of rebuilding them later. |
| Model and training procedure | The trained model cannot be separated from the platform, or nobody other than the provider can repeat the training. | Agree that you will receive the model as a container or in an open format, and that the training is documented so that third parties can follow it. |
| Operational knowledge held by the provider or supplier | Only the provider or supplier knows how the model is monitored and retrained and what to do in the event of an incident. | Agree on documentation, a runbook for incidents, and a training session as part of the agreed services. |
| Contract term and licensing model | A long minimum term binds you even when the solution no longer fits. With prices per user or per request, costs rise with usage. | Before you sign, work through the contract term, the termination notice period, and how prices may change. Find out which components require a license and how the provider supports a switch. |
Open standards exist for containers and for model formats. A container bundles the model with its runtime environment and its dependencies. The Open Container Initiative was founded under the auspices of the Linux Foundation and maintains specifications for the format, the runtime, and the distribution of containers [2].
ONNX is an open format for machine learning models. It is intended to allow a model to be used with different tools and runtimes [3]. You have to check case by case whether a particular model can be converted to this format and still delivers the same results afterward.
When a Commitment to One Provider Is Acceptable
A commitment to one provider is not harmful in itself. A company that rents a platform operated by the provider leaves operation and maintenance to the provider and accepts dependence in return. For a company without an operations team, such a platform can be the right choice.
A commitment is acceptable if you have estimated the cost of a later switch, the benefit of the platform justifies that cost in your view, and the contract sets out the terms of exit. A risk arises when nobody in the company knows what a switch would cost, or when the provider can change prices and terms and a switch is no longer economically feasible.
For GPUs, we calculated the cost of running models on owned hardware in four scenarios. In each of these scenarios, the rented EU cloud was cheaper than owned hardware. For the smallest server in our calculation, the staff needed to run it cost more than four times the monthly depreciation.
The full calculation, based on 110 dated price sources (prices as of July 2026), is in our article Sovereign AI, honestly priced: what on-premises really costs.What Switching Providers Costs: A Model Calculation of Switching Costs
You can estimate the cost of a switch before you sign a contract. The following model calculation shows the method using a fictitious case. It uses person-days and monthly fees so that you can apply your own rates. All input values are arbitrary assumptions and are not based on experience.
Sample calculation for moving a demand forecast to another provider (Model calculation). A company runs a demand forecast on a provider's platform and pays a monthly fee for the service. It switches to another provider. The work involved in the switch is stated in person-days, regardless of whether the company's own staff or a supplier does the work.
| Item | Assumption |
|---|---|
| Export the data and convert it to a different format | 15 person-days |
| Rebuild functions that existed only on the previous platform | 40 person-days |
| Retrain and test the model because it cannot be transferred | 25 person-days |
| Onboard a new operations team | 10 person-days |
| Parallel operation of the old and the new solution | 3 monthly fees |
| Remaining term of the old contract after shutdown | 4 monthly fees |
| Total | 90 person-days and 7 monthly fees |
The total is 90 person-days and 7 monthly fees. To get the amount in euros, multiply the person-days by your rate per person-day and the months by your monthly fee. The calculation does not include a charge from the previous provider for the switch. The running costs of the new solution are also left out because they are not costs of the switch.
Five of the six items each correspond to one of the five areas in the table above. If the model can be transferred as a container and the solution uses no functions that exist only on the previous platform, the second and third items drop out of the model calculation. That leaves 25 person-days and 7 monthly fees. For each month by which the remaining term is shorter, the total falls by one more monthly fee.
Only your contract and your solution show how large the items are in your case. For that reason, enter your own values into the calculation.
What the EU Data Act Says About Switching Providers
The EU regulation known as the Data Act contains a separate chapter on switching between data processing services [1]. The regulation uses the term data processing service for a service that provides on-demand network access to computing resources from a shared pool. According to its recitals, this includes the three common cloud models, namely infrastructure, platform, and software as a service (IaaS, PaaS, and SaaS). The regulation has applied since September 12, 2025.
Providers of data processing services must set out the customer's rights in relation to switching in a written contract. The customer can switch to a different provider or to on-premises ICT infrastructure. The contract must state a maximum notice period for initiating the switching process, and this period may not exceed two months.
From January 12, 2027, providers may no longer impose switching charges for the switching process. Until that date, only reduced switching charges are permitted, and these may not exceed the provider's costs that are directly linked to the switching process. The regulation does not count standard service fees or early termination penalties as switching charges.
The regulation exempts certain services from some of the switching obligations, including the rules on switching charges. These include services in which the majority of main features has been custom-built for an individual customer and which are not offered at broad commercial scale.
Switching charges within the meaning of the regulation are charges imposed by the previous provider. The work inside the customer's own company for transfer, rebuilding, and parallel operation does not count as a switching charge. That work is covered by the model calculation in the preceding section.
This section reflects the content of the regulation as of November 15, 2026, and is not legal advice. In November 2025, the European Commission proposed amendments to the regulation that also affect the chapter on switching and had not been adopted as of that date [4]. You should seek a legal review of whether and how the rules apply to your contract.
Avoiding Vendor Lock-in: How myBytes Works
You remain dependent on a supplier if the supplier is the only one who knows how the model was built and how to run it. That is why we establish at the start of a project what your team needs to receive at handover so that it can continue working without us. You can check how we work against the following four points.
The model runs in a container
The finished model runs in a container, so it can be operated unchanged in a different environment, for example with another provider or in your own data center.
Open-source tools are enough at first
For versioning and delivery, open-source tools are enough as long as a person decides when a model is retrained. A licensed platform is only worth considering once many models and several teams share the same delivery pipeline.
The model and the documentation belong to you
The model we train on your data remains your property. The documentation also belongs to you. It describes what the model does and which data it needs.
Your team can take over operations
We treat the handover to your team as a separate project step. It includes the documentation, a runbook for incidents, and a training session. The goal is for your team to run the model itself afterward.
Until your team takes over operations, we help run the model. We agree with you case by case whether and for how long we continue beyond that point.
We work where your data is stored and, on request, entirely on-premises in your infrastructure.
Frequently Asked Questions About Vendor Lock-in
What is vendor lock-in?
Vendor lock-in is a dependence on a provider that makes switching so expensive or laborious that it does not happen. In AI and data projects, it can arise in five areas. These are data formats, functions available on only one platform, the model and its training, operational knowledge at the provider or supplier, and contract term and licensing model.
What is the lock-in effect?
The lock-in effect is the impact that high switching costs have on the customer. The more expensive a switch would be, the weaker the incentive to choose a different provider. In AI and data projects, dependence on the provider usually grows with the length of use, because the provider holds more data and more interfaces to the company's own systems are added.
How can you avoid vendor lock-in in AI?
Before signing, find out in which format you will get your data back and whether the trained model runs without the provider's platform, for example as a container. Documentation and training should be part of the agreed services. Also work through the term, the termination notice period, and how prices may change, and find out which components require a license.
Is vendor lock-in always a disadvantage?
No. A commitment to one provider is not harmful in itself. A platform operated by the provider relieves your company of operation and maintenance and can be the right choice for a company without an operations team. The commitment is acceptable if you have estimated the cost of a later switch, the benefit justifies that cost, and the contract sets out the terms of exit.
What does switching providers cost?
The cost depends on the case and can be estimated before signing. Our model calculation, which uses arbitrary assumptions, arrives at 90 person-days and 7 monthly fees. If the model can be transferred as a container and no platform functions need rebuilding, 25 person-days and 7 monthly fees remain in this calculation. You get the amount in euros by applying your own rates.
What does the EU Data Act say about switching providers?
Regulation (EU) 2023/2854 has applied since September 12, 2025, and governs switching between data processing services. The maximum notice period for initiating the switching process may not exceed two months. From January 12, 2027, providers may no longer impose switching charges for the switching process. This information reflects the status as of November 15, 2026, and is not legal advice.
Further Pages on Sovereignty, Operations, and Costs
Digital sovereignty in AI: four checks for any offer
Custom AI model or rented off-the-shelf software
MLOps explained: maturity levels, tools, and hosting location in phase 6
On-premises or cloud: the cost calculation from 110 price sources
Sources
[1] Regulation (EU) 2023/2854 of the European Parliament and of the Council of 13 December 2023 on harmonised rules on fair access to and use of data and amending Regulation (EU) 2017/2394 and Directive (EU) 2020/1828 (Data Act), OJ L, 2023/2854, 22.12.2023. Chapter VI (Articles 23 to 31), Article 2, points 8 and 36, Article 50, recitals 81, 88, and 90. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32023R2854 (accessed November 15, 2026)
[2] Open Container Initiative: About the Open Container Initiative. https://opencontainers.org/about/overview/ (accessed November 15, 2026)
[3] ONNX: Open Neural Network Exchange. https://onnx.ai/ (accessed November 15, 2026)
[4] European Commission: Proposal for a Regulation of the European Parliament and of the Council amending, among others, Regulation (EU) 2023/2854 as regards the simplification of the digital legislative framework (Digital Omnibus), COM(2025) 837 final, 19.11.2025, procedure 2025/0360(COD). https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52025PC0837 (accessed November 15, 2026)
Let us talk about your data
Free 30-minute initial call. Briefly describe your dataset and your plan; optionally you can state a preferred date.
These services do not replace legal data protection advice.